Results visualizer — mlperf-viz

Renders a submission folder as a local browser dashboard — the Pareto Explorer — so you can look at your curves before submitting, or browse published results. It reads the same layout the submission checker expects. It does not validate anything: use the checker for that.

Ships as its own package, mlperf-viz, from mlcommons/mlperf-viz. Nothing leaves your machine — no account, no token, no upload.

Install

pip install mlperf-viz
mlperf-viz -V

Usage

mlperf-viz [PATH] [options]

PATH is the submission tree root and defaults to the current directory. It may be a whole results repository, a single organisation's directory, or one submission.

Flag	Description
--version {1.0,0.7}	Submission format to read. Default 1.0
--host HOST	Interface to bind. Default 127.0.0.1
--port PORT	Preferred port; the next free one is used if taken. Default 8000
--no-browser	Do not open a browser window
--export FILE	Write the dashboard dataset as JSON and exit
--build DIR	Write a self-contained copy of the site plus data to DIR and exit
--strict	Exit non-zero if any structural warning was raised
-V, --cli-version	Print the tool's own version and exit

--version is the submission format, not the tool version

--version 0.7 reads the v0.7 layout. -V prints which mlperf-viz you have installed.

Exit codes: 0 on success · non-zero under --strict when a structural warning was raised. Without --strict, missing or unreadable files are reported as warnings and skipped, so a partial submission still renders.

Formats

	v1.0 (default)	v0.7 (--version 0.7)
Anchored on	results/<System>/<Model>/r<N>/ point directories	systems/*.json
System description	system_desc.json, one per point	systems/<id>.json, one per model directory
Point config	point.yaml, inside the point	points/point_<N>.yaml
Metrics	result_summary.json	run_metadata.json
Accuracy	accuracy_results.json	on the system description
Dashboard entry	<Org>-<submission_id>-<System>-<Model>	<Org>-<System>-<Model>

One <System>/<Model> pair is one Pareto curve and one dashboard entry, so a submission covering two models shows as two entries. The v1.0 layout itself is described in Submission package layout.

For v1.0, system_tps, tps_per_user and qps are derived from the raw stat blocks when the summary does not already report them:

system_tps   = output_sequence_lengths.total / (duration_ns / 1e9)
tps_per_user = 1000 / tpot_p90_ms
qps          = n_samples_completed / (duration_ns / 1e9)

v0.7's run_metadata.json already carries all three.

Example: your own submission (v1.0)

Point it at the same folder you pass to submission-checker check:

mlperf-viz /path/to/submission

It prints how many submissions and runs it found, starts a local server and opens the Pareto Explorer. Add --no-browser on a remote machine and forward the port.

Example: v0.7 published results

The v0.7 results are published in the public repository mlcommons/endpoints_results_v0.7, in the published-results variant of the v0.7 layout (<Org>/<System>/<Model>/).

git clone https://github.com/mlcommons/endpoints_results_v0.7.git
mlperf-viz endpoints_results_v0.7 --version 0.7

Found 10 submission(s) and 100 run(s) in v0.7 format.

To look at one organisation, point at its directory:

mlperf-viz endpoints_results_v0.7/Krai --version 0.7

To share a snapshot without anyone installing anything, build a static copy and host the directory anywhere that serves files:

mlperf-viz endpoints_results_v0.7 --version 0.7 --build v0.7-site

To use the data programmatically instead, --export writes the dashboard dataset — top-level submissions and runs arrays — as JSON:

mlperf-viz endpoints_results_v0.7 --version 0.7 --export v0.7.json

Caveats

• The format is chosen by the flag, not detected. Reading a v0.7 tree without --version 0.7 finds no submissions; the tool prints the layout it expected and names the flag.
• v0.7 has a second layout — the submission-checker variant, <Org>/systems/<system_desc_id>.json with pareto/<system_desc_id>/<model>/{points,results}/ — which --version 0.7 also reads.
• Entries are identified by path, not by the submission_id field, which is frequently absent, empty, or shared between distinct submissions.
• One accelerator per node. A v1.0 node type that discloses more than one accelerator configuration is shown with the first, and a warning says how many were omitted.
• Fonts and analytics load from Google when online; offline the dashboard falls back to system fonts.

Last verified against: mlcommons/mlperf-viz@cli-v1.1.0 (654dca5) with mlperf-viz 1.1.0 from PyPI, and mlcommons/endpoints_results_v0.7@main, 2026-10-02.